CVPR 2026 Papers — Page 36
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
TextPecker: Rewarding Structural Anomaly Quantification for Enhancing Visual Text Rendering
Hanshen Zhu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
GenerationAnomaly DetectionReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposes TextPecker, which employs a structure-aware anomaly detection-based reward strategy to perform reinforcement learning on text generation models, thereby enhancing the structural integrity and semantic consistency of visual text rendering.
TEXTRIX: Latent Attribute Grid for Native Texture Generation and Beyond
Yifei Zeng (Nanjing University), Yao Yao (Nanjing University)
SegmentationGenerationTransformerDiffusion modelFlow-based ModelRectified FlowAuto EncoderImageMesh
🎯 What it does: Proposes a native 3D generative framework (TEXTRIX) that directly generates 3D textures and semantic labels on sparse voxel grids, achieving high-precision 3D segmentation within the same architecture.
Texvent: Asynchronous Event Data Simulation via Text Prompt
Ruofei Wang (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
GenerationData SynthesisDepth EstimationTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: Texvent proposes a training-free text-to-event (T2E) simulation framework that directly generates asynchronous, high-resolution event data from text prompts using multimodal large language models.
TF-CADE: Foreground-Concentrated Text-Video Alignment for Zero-Shot Temporal Action Detection
Yearang Lee (Korea University), Seong-Whan Lee (Korea University)
RecognitionTransformerVision Language ModelVideoText
🎯 What it does: Propose the TF-CADE framework, achieving zero-shot temporal action detection through foreground-focused text-video alignment;
TF-SSD: A Strong Pipeline via Synergic Mask Filter for Training-free Co-salient Object Detection
Zhijin He (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Xi'an Jiaotong-Liverpool University)
Object DetectionTransformerImageBenchmark
🎯 What it does: Proposed a new no-training method called TF-SSD for cooperative significant object detection (CoSOD), generating and filtering masks through the synergistic effects of SAM and DINO.
TGSFormer: Scalable Temporal Gaussian Splatting for Embodied Semantic Scene Completion
Rui Qian (Nanyang Technological University), Lihua Xie (Nanyang Technological University)
SegmentationDepth EstimationAutonomous DrivingComputational EfficiencyRepresentation LearningTransformerGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes TGSFormer, an expandable time Gaussian splatting framework for completing the construction of 3D semantic scenes from continuous first-person perspective observations.
TGT: Text-Grounded Trajectories for Locally Controlled Video Generation
Guofeng Zhang (ByteDance), Chongyang Ma
GenerationTransformerVision Language ModelDiffusion modelFlow-based ModelVideoText
🎯 What it does: Propose a controllable video generation framework based on text-localized trajectory (TGT), allowing users to precisely control the appearance and motion of multiple objects in videos through point trajectories and corresponding local text descriptions.
TGTrack: Temporal Generative Learning for Unified Single Object Tracking
Wanting Geng, Huchuan Lu (Dalian University Of Technology)
Object TrackingGenerationTransformerVideoMultimodality
🎯 What it does: Propose TGTrack, a unified multi-modal single-target tracking framework that explicitly guides the model to capture the evolution of targets and scenes over time through temporal generation learning tasks, achieving efficient tracking across RGB, depth, thermal imaging, event streams, and natural language descriptions.
The Blind Spot of Adaptation: Quantifying and Mitigating Forgetting in Fine-tuned Driving Models
Runhao Mao (Shanghai Jiao Tong University), Zhipeng Zhang (Shanghai Jiao Tong University)
Autonomous DrivingTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsVision Language ModelMultimodalityBenchmark
🎯 What it does: Systematically studied catastrophic forgetting of vision-language models (VLM) during fine-tuning for autonomous driving, and proposed a benchmark called 'Fidelity Driving Bench' based on large-scale long-tail scenarios.
The Coherence Trap: When MLLM-Crafted Narratives Exploit Manipulated Visual Contexts
Yuchen Zhang (School of Software Engineering, Xi'an Jiaotong University), Zhedong Zheng (University of Macau)
ClassificationData SynthesisAnomaly DetectionLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Constructed a large-scale semantically aligned multimodal fake news dataset (MDSM) and proposed a framework (AMD) that utilizes multimodal large language models (MLLM) for generating and detecting forged text.
The Consistency Critic: Correcting Inconsistencies in Generated Images via Reference-Guided Attentive Alignment
Ziheng Ouyang (Nankai University), Mike Zheng Shou (National University of Singapore)
Image HarmonizationRestorationTransformerSupervised Fine-TuningAgentic AIPrompt EngineeringVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: Propose ImageCritic, a post-processing framework based on reference images, which corrects fine-grained inconsistencies in generated images using attention alignment loss and a detail encoder, and constructs a corresponding reference-degraded-target dataset;
The Devil is in Attention Sharing: Improving Complex Non-rigid Image Editing Faithfulness via Attention Synergy
Zhuo Chen, Wen Li
TransformerDiffusion modelImageTextBenchmark
🎯 What it does: Propose a training-free complex non-rigid image editing method, SynPS, which utilizes an attention sharing mechanism to achieve precise semantic transformation between the source image and the target text instruction.
The Devil Is in Gradient Entanglement: Energy-Aware Gradient Coordinator for Robust Generalized Category Discovery
Haiyang Zheng (University of Trento), Zhun Zhong (Hefei University of Technology)
ClassificationRepresentation LearningTransformerContrastive LearningImageBenchmark
🎯 What it does: This paper proposes the Energy-Aware Gradient Coordinator (EAGC) module in the Generalized Category Discovery task to alleviate gradient entanglement issues.
The devil is in the details: Enhancing Video Virtual Try-On via Keyframe-Driven Details Injection
Qingdong He (Tencent Youtu Lab), Yabiao Wang (Zhejiang University)
Image TranslationGenerationPose EstimationTransformerPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderImageVideo
🎯 What it does: Proposed a keyframe-driven detail injection framework called KeyTailor based on Diffusion Transformer to enhance detail fidelity and background consistency in video virtual try-on.
The Drift Kernel: Why Diffusion Models Change Even When Told Not To
Gokul Srinath Seetha Ram (Independent Researcher), Rashmi Elavazhagan (Independent Researcher)
GenerationDiffusion modelImageBenchmark
🎯 What it does: This paper investigates how diffusion models generate image drift even under 'do nothing' or no-edit prompts, and proposes Drift Kernel to quantify this drift;
The Geometry of Robustness: Optimizing Loss Landscape Curvature and Feature Manifold Alignment for Robust Finetuning of Vision-Language Models
Shivang Chopra (Georgia Institute of Technology), Zsolt Kira (Georgia Institute of Technology)
OptimizationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a unified VLM fine-tuning framework called GRACE to achieve a balance among in-distribution (ID), out-of-distribution (OOD), and adversarial robustness.
The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation
Guannan Lai (Nanjing University), Han-Jia Ye (Nanjing University)
ClassificationSegmentationDomain AdaptationSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Proposed a source-free continuous test-time adaptation framework called GOLD, which utilizes a low-rank golden subspace to perform lightweight feature updates, addressing the trade-off between efficiency and generalization.
The Image as Its Own Reward: Reinforcement Learning with Adversarial Reward for Image Generation
Weijia Mao (Show Lab, National University of Singapore), Mike Zheng Shou (Show Lab, National University of Singapore)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelContrastive LearningImageText
🎯 What it does: This study proposes Adv-GRPO, a reinforcement learning framework based on adversarial rewards, aimed at improving the quality of text-to-image generation.
The Invisible Gorilla Effect in Out-of-distribution Detection
Harry Anthony (University of Oxford), Konstantinos Kamnitsas (University of Oxford)
Anomaly DetectionTransformerImageBiomedical DataBenchmark
🎯 What it does: Evaluated 40 OOD detection methods through large-scale experiments and found that models better detect OOD artefacts that are visually similar to the ROI (i.e., the Invisible Gorilla Effect).
The LLM Bottleneck: Why Open-Source Vision LLMs Struggle with Hierarchical Visual Recognition
Yuwen Tan, Boqing Gong (Boston University)
ClassificationRecognitionLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: This paper constructs approximately one million four-option visual question answering (VQA) tasks to systematically evaluate the performance of publicly available visual large language models (VLLMs) in hierarchical visual recognition, revealing that their core bottleneck lies in the lack of taxonomic knowledge about the visual world in the underlying language models;
The Midas Touch for Metric Depth
Yu Ma (Tongji University), Rui Fan (Tongji University)
Depth EstimationKnowledge DistillationTransformerImage
🎯 What it does: Propose a training-free framework called Midas Touch for Depth (MTD), which converts relative depth into metric depth through sparse 3D seeds, and performs scale recovery at the graph level and pixel-level refinement.
The Missing GAP: From Solving Square Jigsaw Puzzles to Handling Real World Archaeological Fragments
Ofir Itzhak Shahar (Ben Gurion University Negev), Ohad Ben-Shahar (Ben Gurion University Negev)
RestorationData SynthesisTransformerFlow-based ModelAuto EncoderImage
🎯 What it does: Created the GAP dataset (containing 20,000 irregularly eroded puzzle pieces generated from real archaeological fragment distributions) and proposed the PuzzleFlow framework to solve non-square, irregularly shaped, and severely edge-eroded puzzle problems.
The Missing Point in Vision Transformers for Universal Image Segmentation
Sajjad Shahabodini (Concordia University), Arash Mohammadi (Concordia University)
SegmentationTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose a two-stage segmentation framework ViT-P: first generate masks using a class-agnostic mask generator, then use a Vision Transformer-based point classifier to precisely classify the masks, thereby significantly improving segmentation accuracy.
THE MORE, THE MERRIER: CONTRASTIVE FUSION FOR HIGHER-ORDER MULTIMODAL ALIGNMENT
Stefanos Koutoupis (Foundation for Research and Technology-Hellas), Grigorios Tsagkatakis (Foundation for Research and Technology-Hellas)
ClassificationRetrievalRepresentation LearningContrastive LearningImageTextMultimodalityAudio
🎯 What it does: Propose Contrastive Fusion (ConFu), a contrastive learning framework that jointly aligns single-modal representations with their fused representations.
The Power of Decaying Steps: Enhancing Attack Stability and Transferability for Sign-based Optimizers
Wei Tao (National University of Defense and Technology), Qing Tao (Hefei Institute of Technology)
ClassificationRetrievalAdversarial AttackConvolutional Neural NetworkTransformerImageMultimodality
🎯 What it does: This paper improves upon adversarial attack algorithms based on sign gradients by proposing the MDCS (Monotonically Decreasing Coordinate Step Size) strategy, which is embedded into classical methods such as I-FGSM and MI-FGSM, significantly enhancing the stability and transferability of attacks.
The Power of Prior: Training-Free Open-Vocabulary Semantic Segmentation with LLaVA
Bingfeng Zhang (China University Of Petroleum East China), Jimin Xiao (Xjtlu)
SegmentationLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: This paper proposes a fully training-free open-source lexical semantic segmentation method called FSeg-LLaVA, which leverages the internal prior knowledge of the multimodal large language model LLaVA. It identifies foreground classes through a question-answering pipeline, extracts visual-textual responses, refines them through prototyping, and generates point/box prompts for SAM to produce the final mask.
The Road Less Seen: Segment Exploration for Weakly Supervised Video Anomaly Detection
Anusha Acharya (Rochester Institute of Technology), Xumin Liu (Rochester Institute of Technology)
Anomaly DetectionVision Language ModelVideo
🎯 What it does: This paper proposes a dual exploration strategy for weakly supervised video anomaly detection, leveraging temporal clustering and uncertainty sampling to more comprehensively select abnormal segments from positive videos to enhance detection performance.
The SA-FARI Dataset: Segment Anything in Footage of Animals for Recognition and Identification
Dante Wasmuht (Conservation X Labs), Didac Suris (Meta)
RecognitionObject DetectionObject TrackingSegmentationSupervised Fine-TuningPrompt EngineeringVideoBenchmark
🎯 What it does: Constructed and publicly released SA-FARI, the first largest-scale and most species-diverse multi-animal tracking dataset covering 99 wildlife species, and conducted benchmarking tests on multiple visual models based on this dataset.
The Surprising Effectiveness of Noise Pretraining for Implicit Neural Representations
Kushal Vyas (Rice University), Guha Balakrishnan (Rice University)
Representation LearningImageVideoMagnetic Resonance Imaging
🎯 What it does: Explore the effectiveness of noise pretraining for initializing implicit neural representations (INR) and propose the SNP (STRIANER Noise Pretraining) method
The Universal Normal Embedding
Chen Tasker (Israel Institute of Technology), Guy Gilboa (Israel Institute of Technology)
ClassificationGenerationTransformerDiffusion modelContrastive LearningImage
🎯 What it does: Proposed the Universal Normal Embedding (UNE) hypothesis, treating the latent space of generative models and visual encoders as noisy linear projections of the same Gaussian latent space; created the NoiseZoo dataset by pairing DDIM inverted noise with embeddings from multiple encoders (CLIP, OpenCLIP, DINOv3); evaluated the Gaussianity, linear separability, cross-model alignment, and controllable editing via linear directions in the latent space on CelebA.
TherA: Thermal-Aware Visual-Language Prompting for Controllable RGB-to-Thermal Infrared Translation
Dong-Guw Lee (Seoul National University), Ayoung Kim (Seoul National University)
Image TranslationTransformerPrompt EngineeringVision Language ModelDiffusion modelMultimodality
🎯 What it does: Proposed a controllable RGB-to-thermal infrared (TIR) image translation framework called TherA, capable of generating TIR images that comply with thermal physics laws, and supporting control via text and reference images
Thermal Diffusion Matters: Infrared Spatial-Temporal Video Super-Resolution through Heat Conduction Priors
Mingxuan Zhou (Beijing Institute of Technology), Shuigen Wang (Iray Technology Co., Ltd)
Super ResolutionDiffusion modelVideoPhysics Related
🎯 What it does: Proposes a THERIS framework for infrared spatiotemporal video super-resolution based on thermal diffusion prior, achieving dual enhancement of spatial and temporal resolution for low-resolution low-frame-rate infrared videos.
Thermal is Always Wild: Characterizing and Addressing Challenges in Thermal-Only Novel View Synthesis
M. Kerem Aydin (Northwestern University), Emma Alexander (Northwestern University)
GenerationData SynthesisGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes a lightweight preprocessing and 3D Gaussian Splatting reconstruction framework specifically tailored for thermal imaging, enabling high-quality novel view synthesis using only thermal images.
Thermal-Det: Language-Guided Cross-Modal Distillation for Open-Vocabulary Thermal Object Detection
Yasiru Ranasinghe (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
Object DetectionData SynthesisKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes Thermal-Det, a zero-annotation thermal imaging open-vocabulary detection framework, achieving zero-shot object detection on thermal images through synthetic thermal data generation, cross-modal distillation, and thermal-text alignment.
Thermally Activated Dual-Modal Adversarial Clothing against AI Surveillance Systems
Jiahuan Long (Defense Innovation Institute, Chinese Academy of Military Science), Wen Yao (Defense Innovation Institute, Chinese Academy of Military Science)
Object DetectionAdversarial AttackImageMultimodality
🎯 What it does: Developed a thermally activated wearable adversarial garment that achieves dual-modal interference against AI surveillance in both visible and infrared modalities through thermochromic microcapsules and heating elements.
Think 360deg: Beyond Depth: Evaluating the Width-centric Reasoning Capability of MLLMs
Mingrui Chen (University of Chinese Academy of Sciences), Ran He (University of Chinese Academy of Sciences)
Large Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes a width-oriented multimodal reasoning benchmark called Think360, collecting, screening, and annotating over 1200 multimodal cases, and using a tree-of-thought evaluation method to conduct fine-grained assessments of the width and depth of MLLMs.
Think Before You Drive: World Model-Inspired Multimodal Grounding
Haicheng Liao (University of Macau), Zhenning Li (University of Macau)
Depth EstimationAutonomous DrivingGraph Neural NetworkTransformerVision Language ModelWorld ModelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the ThinkDeeper framework, which predicts future hidden states through a spatial-aware world model and achieves visual grounding in autonomous driving scenarios using a multi-modal hypergraph decoder.
Think Visually, Reason Textually: Vision-Language Synergy in Abstract Reasoning
Beichen Zhang (Chinese University of Hong Kong), Jiaqi Wang (Shanghai AI Laboratory)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper studies the application of joint vision and text reasoning to the ARC-AGI abstract reasoning task, proposing two strategies: Vision-Language Synergy Reasoning (VLSR) and Modality-Switch Self-Correction (MSSC), which allocate the optimal modality during rule summarization and rule execution stages respectively;
Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views
Zhangquan Chen (Tsinghua University), Ruqi Huang (Tsinghua University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: Developed the 3DThinker framework, enabling vision-language models to spontaneously generate 3D latent layers during reasoning, thereby achieving the ability to 'think in 3D' without requiring external 3D priors or manual annotations.
Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models
Jialiang Zhang (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)
Computational EfficiencyTransformerSupervised Fine-TuningVision Language ModelVideoTextChain-of-Thought
🎯 What it does: Propose the Think-as-You-See (TaYS) framework to enable real-time incremental inference of large-scale vision-language models in video streams.
Think-Then-Generate: Structural Chain-of-Thought Reasoning for Consistent 3D Generation
Xinyue Liu (Chinese Academy of Sciences), Huaibo Huang (Chinese Academy of Sciences)
GenerationData SynthesisLarge Language ModelVision Language ModelDiffusion modelScore-based ModelImageTextMultimodalityPoint CloudMeshChain-of-Thought
🎯 What it does: Propose the Thoughtful3D framework, which significantly enhances multi-view consistency and reduces hallucinations by planning, reflecting, and correcting throughout the 3D generation process through chain-of-thought (CoT).
Think, Then Verify: A Hypothesis-Verification Multi-Agent Framework for Long Video Understanding
Zheng Wang (Zhejiang University of Technology), Cong Bai (Zhejiang University of Technology)
RetrievalExplainability and InterpretabilityComputational EfficiencyLarge Language ModelAgentic AIVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the VideoHV-Agent multi-agent framework, which decomposes the long-video question answering task into four steps: hypothesis generation, clue extraction, evidence retrieval, and answer integration. It first locates potential answers using video summaries, then precisely retrieves fine-grained evidence from the minimal time window through a hypothesis-verification process.
ThinkGen: Generalized Thinking for Visual Generation
Siyu Jiao (Beijing Jiaotong University), Yunchao Wei (Beijing Jiaotong University)
GenerationTransformerReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes ThinkGen, a visual generation framework that guides diffusion models to generate high-quality images through chain-of-thought (CoT) mechanisms of multimodal large language models.
Thinking Beyond Labels: Vocabulary-Free Fine-Grained Recognition using Reasoning-Augmented LMMs
Dmitry Demidov (Mohamed bin Zayed University of Artificial Intelligence), Rao Anwer (Mohamed bin Zayed University of Artificial Intelligence)
RecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: Propose an automated, vocabulary-free fine-grained recognition framework based on a reasoning-enhanced large multimodal model (FiNDR).
Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models
Keuntae Kim (Hanyang University), Yong Suk Choi (Hanyang University)
Explainability and InterpretabilityComputational EfficiencyDiffusion modelMultimodalityChain-of-Thought
🎯 What it does: Analyze the behavior of diffusion-based multi-modal large language models (dMLLM) during the Chain-of-Thought (CoT) reasoning process, revealing their tendency to generate early answers and underutilize visual information, and propose two zero-training inference-time techniques—Position & Step Penalty (PSP) and Visual Reasoning Guidance (VRG)—to improve reasoning quality and speed.
Thinking in 360deg: Humanoid Visual Search in the Wild
Heyang Yu (New York University), Yiming Li (New York University)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelVideoMultimodalityBenchmark
🎯 What it does: Propose Humanoid Visual Search (HVS), enabling AI to perform target and path search in 360° panoramic images through active head rotation, simulating the human visual search process.
Thinking in Dynamics: How Multimodal Large Language Models Perceive, Track, and Reason Dynamics in Physical 4D World
Yuzhi Huang (Xiamen University), Zhi Wang (Tsinghua University)
Object TrackingTransformerLarge Language ModelPrompt EngineeringVideoTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the Dyn-Bench benchmark to evaluate the spatiotemporal perception, tracking, and reasoning capabilities of multimodal large language models (MLLMs) in 4D dynamic worlds; simultaneously studied and validated the effectiveness of structured text and vision-guided methods (ST-TCM and Mask-Guided Fusion) in enhancing dynamic reasoning and target localization.
Thinking in Uncertainty: Mitigating Hallucinations in MLRMs with Latent Entropy-Aware Decoding
Zhongxing Xu (Monash University), Zongyuan Ge (Monash University)
Explainability and InterpretabilityLarge Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose an adaptive decoding strategy called LEAD based on token entropy to alleviate hallucinations in multi-modal large-scale reasoning models (MLRMs).
Thinking with Drafts: Speculative Temporal Reasoning for Efficient Long Video Understanding
Pengfei Hu (Alibaba Group Holding Limited), Xiaodan Liang (Alibaba Group Holding Limited)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodality
🎯 What it does: Propose a dual-model collaborative framework called SpecTemp, which uses a lightweight Draft MLLM to quickly select keyframes and a heavyweight Target MLLM for temporal reasoning and verification, thereby achieving efficient long-video understanding.
Thinking with Frames: Generative Video Distortion Evaluation via Frame Reward Model
Yuan Wang (University Of Science And Technology Of China), Xiang Wang (Kling Team Kuaishou Technology)
Anomaly DetectionSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextBenchmarkChain-of-Thought
🎯 What it does: Proposes REACT — a frame-level reward model for evaluating structural deformations in generated videos; constructs a detailed classification of structural deformations along with human preference and attribution label datasets, and expands training samples through an efficient Chain-of-Thought synthesis pipeline; adopts a two-stage training (supervised fine-tuning + GRPO reinforcement learning) on Qwen2.5-VL-7B with a dynamic frame sampling mechanism, ultimately providing a specialized REACT-Bench benchmark.
Thinking with Programming Vision: Towards a Unified View for Thinking with Images
Zirun Guo (Zhejiang University), Tao Jin (Zhejiang University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposed the CodeVision framework, enabling multimodal large language models to call arbitrary image tools through code generation, achieving more flexible and scalable image thinking capabilities.
Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm
Jingqi Tong (Shanghai Innovation Institution), Xipeng Qiu (Fudan University)
GenerationPrompt EngineeringVision Language ModelDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: Propose the 'Thinking with Video' paradigm, using video generation models for multimodal reasoning and constructing the VideoThinkBench benchmark.
Thinking With Videos: Multimodal Tool-Augmented Reinforcement Learning for Long Video Reasoning
Haoji Zhang (Tsinghua University), Yansong Tang (Tsinghua University)
Large Language ModelReinforcement LearningAgentic AIVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed an end-to-end video tool-enhanced reasoning framework called VITAL, which enables multi-modal large language models to actively request and utilize new video frames through a visual toolbox, generating multi-modal chain reasoning processes; constructed a large-scale multi-task video reasoning dataset MTVR-CoT-72k and MTVR-RL-110k, and used the difficulty-aware GRPO algorithm for reinforcement learning based on these datasets.
Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation
Ziyu Guo (CUHK), Pheng-Ann Heng (Meituan)
GenerationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes a framework called Thinking-while-Generating (TWIG), enabling real-time interaction between text reasoning and image generation during the visual generation process, providing immediate guidance and correction at each step of generation.
ThinkingViT: Matryoshka Thinking Vision Transformer for Elastic Inference
Ali Hojjat (Kiel University), Olaf Landsiedel (Kiel University)
ClassificationSegmentationComputational EfficiencyTransformerImage
🎯 What it does: Proposed a nested Vision Transformer architecture called ThinkingViT, which utilizes progressively activating more attention heads combined with the Token Recycling mechanism to achieve input-adaptive inference.
Through the Frequency Lens: Cross-Domain Generalisable Gaze Estimation with Adaptive Modulation
Yang Xu (Beihang University), Feng Lu (Beihang University)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: Propose a cross-domain generalizable gaze estimation framework FGAL, combining the Adaptive Interference Suppression Module (AISM) and the Spectral Diversification Module (SDM);
TIACam: Text-Anchored Invariant Feature Learning with Auto-Augmentation for Camera-Robust Zero-Watermarking
Abdullah Tanvir (University of Nebraska Omaha), Xin Zhong (University of Nebraska Omaha)
TransformerVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: Propose the TIACam framework to achieve a camera-robust zero-watermark system.
TIGER: A Unified Framework for Time, Images and Geo-location Retrieval
David G. Shatwell (University of Central Florida), Mubarak Shah (University of Central Florida)
RetrievalTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Propose a unified multimodal Transformer framework named TIGeR, which models images, geolocation, and time simultaneously, enabling location- and time-based image retrieval, time prediction, and geolocation.
TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation
Yiheng Dong (Huazhong University of Science and Technology), Xin Yang (Huazhong University of Science and Technology)
GenerationTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityElectronic Health Records
🎯 What it does: Proposed the TIM model, utilizing a two-stage framework with temporal decoupling and mutual iterative refinement to generate longitudinal X-ray imaging reports.
Time Blindness: Why Video-Language Models Can't See What Humans Can?
Ujjwal Upadhyay (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
Data SynthesisExplainability and InterpretabilitySupervised Fine-TuningOptical FlowVideoTextBenchmarkChain-of-Thought
🎯 What it does: Propose SpookyBench, a benchmark that constructs videos composed of noisy frames where information exists only in the temporal dimension, to evaluate the pure temporal understanding capability of video-language models.
Time Without Time: Pseudo-Temporal Representation for Space-Time Super-Resolution
Hee Min Choi (Samsung Electronics Co., Ltd), Nam Ik Cho (Seoul National University)
Super ResolutionImageVideo
🎯 What it does: Propose a pseudo-temporal spatial reconstruction pre-training framework, which pre-trains a fixed frame rate space-time video super-resolution network by generating pseudo-temporal videos through copying a single image and randomly masking it.
Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution
Tianyi Zhang (Nankai University), Chongyi Li (Nankai University)
Super ResolutionSupervised Fine-TuningDiffusion modelAuto EncoderImage
🎯 What it does: Proposed a time-aware one-step diffusion network called TADSR for real-world image super-resolution, leveraging the prior knowledge of Stable Diffusion and regulating realism and fidelity through time steps.
Time-Specialized Event-Image Alignment for Blur-to-Video Decomposition
Zhijing Sun (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationGenerationTransformerOptical FlowImageVideo
🎯 What it does: Achieving high-frame-rate video decomposition from single-frame motion-blurred images using event cameras and deep learning.
TimeBridge: Self-Supervised Video Representation Learning via Start-End Joint Embedding and In-Between Frame Prediction
Qin Wang (Forschungszentrum Julich GmbH), Kai Krajsek (Forschungszentrum Julich GmbH)
Representation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: This paper proposes a self-supervised video representation learning framework called TimeBridge, which is based on the joint embedding of start and end frames and predicts the intermediate frames.
TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs
Jun Zhang (Nanjing University), Limin Wang (Nanjing University)
RecognitionTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: This paper proposes TimeLens, a systematic study aimed at establishing a baseline for multimodal large language models (MLLMs) with strong video temporal localization capabilities, focusing on data quality and algorithm design.
TimeRipples: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space
Wenxuan Mao, Yu Feng (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelVideoBenchmark
🎯 What it does: This work proposes an acceleration method called TIMERIPPLE, which reuses self-attention calculations in the vDiT model by leveraging spatiotemporal correlations across channels in the video latent space, significantly reducing the computational cost of self-attention and accelerating video generation.
TimeViper: A Hybrid Mamba-Transformer Vision-Language Model for Efficient Long Video Understanding
Boshen Xu (Renmin University of China), Qin Jin (Renmin University of China)
TransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Propose TimeViper, a hybrid Mamba-Transformer vision-language model designed for long video understanding tasks, introducing the TransV module internally to achieve visual token compression, enabling processing of thousands of frames in long videos.
TINA: Text-Free Inversion Attack for Unlearned Text-to-Image Diffusion Models
Qianlong Xiang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
GenerationOptimizationAdversarial AttackDiffusion modelImageText
🎯 What it does: Propose a text-agnostic reverse attack method called TINA, which utilizes DDIM inversion and self-consistent optimization to recover erased concepts in text-to-image diffusion models where text has been erased;
TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment
Bingyi Cao (Google), Andre Araujo (Google)
ClassificationSegmentationDepth EstimationRetrievalVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose TIPSv2, an image-text pre-training framework that integrates contrastive learning and self-supervised learning, focusing on enhancing the alignment between dense images and text.
TiViBench: Benchmarking Think-in-Video Reasoning for Video Generation
Harold Haodong Chen (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Proposed the TiViBench benchmark to evaluate the inference capabilities of image-to-video generation models, and designed the VideoTPO optimization strategy for testing;
TLMA: Mitigating the Impact of Weakly Labeled Information for Video Anomaly Detection
Rong Xu (Beijing Jiaotong University), Liping Jing (Beijing Jiaotong University)
Anomaly DetectionTransformerContrastive LearningVideo
🎯 What it does: Propose the TLMA framework, combining multiple instance learning (MIL) with triplet learning, designing a motion-aware feature enhancement module tailored for weakly labeled information (WLI) to improve localization accuracy in weakly supervised video anomaly detection.
TM-BSN: Triangular-Masked Blind-Spot Network for Real-World Self-Supervised Image Denoising
Junyoung Park (Seoul National University), Nam Ik Cho (Seoul National University)
RestorationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes the Triangular Mask Blind Spot Network (TM-BSN) for self-supervised real sRGB image denoising.
Token Reduction via Local and Global Contexts Optimization for Efficient Video Large Language Models
Jinlong Li (University Of Trento), Nicu Sebe (University Of Trento)
CompressionOptimizationComputational EfficiencyLarge Language ModelVision Language ModelVideo
🎯 What it does: To address the issue of redundant visual tokens in video large language models (VLLM), this paper proposes a training-free token compression method called AOT, which leverages optimal transport (OT) to aggregate information at both intra-frame and inter-frame levels, achieving efficient video token compression.
Token Warping Helps MLLMs Look from Nearby Viewpoints
Phillip Y. Lee (Korea Advanced Institute of Science and Technology), Minhyuk Sung (Korea Advanced Institute of Science and Technology)
Representation LearningTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes the Token Warping method, enabling multi-modal large language models to perform spatial reasoning and image understanding from adjacent perspectives.
TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens
Jiawei Ren (NVIDIA), Zan Gojcic (NVIDIA)
GenerationTransformerGaussian SplattingImageVideo
🎯 What it does: Propose TokenGS, a feedforward reconstruction framework that decouples 3D coordinate regression from pixels through learnable 3D Gaussian tokens, supporting both static and dynamic scenes and enabling token fine-tuning during testing.
TokenHand: Discrete Token Representation for Efficient Hand Mesh Reconstruction
Xinguo He (Technical University of Munich), Rahul Chaudhari (Technical University of Munich)
Pose EstimationTransformerImageMesh
🎯 What it does: Proposed the TokenHand framework, which quantizes 3D hand meshes into discrete tokens and uses a classification model to predict these tokens from single-view images, achieving high-quality and real-time hand mesh reconstruction.
Tokenization Allows Multimodal Large Language Models to Understand, Generate and Edit Architectural Floor Plans
Sizhong Qin, Xinzheng Lu
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelAuto EncoderImageTextMultimodalityBenchmark
🎯 What it does: Proposes HouseMind, a unified multimodal large language model that can simultaneously understand, generate, and edit architectural floor plans through spatial tokenization.
TokenLight: Precise Lighting Control in Images using Attribute Tokens
Sumit Chaturvedi (Yale University), Zhixin Shu (Adobe)
Image TranslationGenerationData SynthesisTransformerDiffusion modelFlow-based ModelAuto EncoderImage
🎯 What it does: TokenLight achieves precise and continuous control of lighting in images by introducing attribute tokens into diffusion transformers, supporting operations such as light intensity, color, diffusion, 3D position, and turning on/off existing scene light sources.
TokenSplat: Token-aligned 3D Gaussian Splatting for Feed-forward Pose-free Reconstruction
Yihui Li (Beihang University), Di Huang (Beihang University)
GenerationPose EstimationTransformerGaussian SplattingImage
🎯 What it does: TokenSplat proposes a forward inference framework that simultaneously accomplishes 3D Gaussian distribution reconstruction and camera pose estimation for pose-free multi-view images, eliminating traditional iterative optimization processes;
TokenTrace: Multi-Concept Attribution through Watermarked Token Recovery
Li Zhang (Adobe Research), Vishal Asnani (Adobe Research)
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Propose TokenTrace, a dual-condition active watermarking framework, to achieve traceability of multiple concepts (object + style) in generated images.
Too Vivid to Be Real? Benchmarking and Calibrating Generative Color Fidelity
Zhengyao Fang (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
GenerationVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: This paper addresses the issue of color authenticity in text-to-image generation by proposing a large-scale color authenticity dataset (CFD), a multi-modal color authenticity metric (CFM), and an untrained dynamic regulator (CFR) to enhance the color authenticity of generated images.
TopoCL: Topological Contrastive Learning for Medical Imaging
Guangyu Meng (University of Notre Dame), Danny Z. Chen (University of Notre Dame)
Representation LearningTransformerMixture of ExpertsContrastive LearningBiomedical Data
🎯 What it does: Propose the TopoCL framework, combining topological information (e.g., connectivity, holes) with traditional contrastive learning to achieve unsupervised medical image representation learning.
TopoHR: Hierarchical Centerline Representation for Cyclic Topology Reasoning in Driving Scenes with Point-to-Instance Relations
Yifeng Bai (Institute of Intelligent Machines, HFIPS, Chinese Academy of Sciences), Haibin Ling (Westlake University)
Autonomous DrivingTransformerImagePoint Cloud
🎯 What it does: Proposed TopoHR, an end-to-end hierarchical centerline detection and topological reasoning framework, achieving mutual enhancement through recursive interaction between detection and reasoning modules;
Topology-aware Feature Propagation for Unsupervised Non-rigid Point Cloud Correspondence
Haozhe Chen (Northwestern Polytechnical University), Jiaqi Yang (Northwestern Polytechnical University)
Pose EstimationOptimizationTransformerPoint Cloud
🎯 What it does: This paper proposes a topology-aware feature propagation module, which is embedded into a coarse-to-fine propagation and optimization pipeline, achieving unsupervised non-rigid point cloud correspondence.
TopoMA: Topology-Guided Multi-Agent Dense RGB 3D Reconstruction via Distributed Inference
Xuanxuan Zhang (Wuhan University), You Li (Wuhan University)
Autonomous DrivingTransformerSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposes TOPOMA, a real-time end-to-end RGB 3D reconstruction framework for multi-agent collaboration, integrating topological skeleton modeling, decentralized loop closure, and topology-guided residual transmission to achieve distributed inference and global consistency;
TopoMesh: High-Fidelity Mesh Autoencoding via Topological Unification
Guan Luo (Tsinghua University), Jianfeng Zhang (ByteDance Seed)
GenerationAuto EncoderMesh
🎯 What it does: Proposes TopoMesh, a VAE that unifies mesh topology through Dual Marching Cubes (DMC), enabling explicit supervision on vertices, faces, and topology to achieve high-fidelity mesh autoencoding.
TopoSlide: Topologically-Informed Histopathology Whole Slide Image Representation Learning
Shahira Abousamra (Stanford University), Sylvia Plevritis (Stanford University)
RetrievalRepresentation LearningTransformerContrastive LearningBiomedical Data
🎯 What it does: Propose TopoSlide, a self-supervised framework for whole-slide image representation learning that leverages persistent homology (topological data analysis) to simultaneously capture local pathological features and global spatial organization;
TouchDream: 3D Object Completion through Imagined Touch
Yuanbo Wang (Dalian University of Technology), Xin Yang (Dalian University of Technology)
GenerationTransformerDiffusion modelPoint Cloud
🎯 What it does: This paper proposes the TouchDream framework, which generates imagined tactile information through diffusion models and completes 3D object point clouds by combining with coarse point clouds.
Toward Diffusible High-Dimensional Latent Spaces: A Frequency Perspective
Bolin Lai (Meta AI), Ishan Misra (Meta AI)
GenerationDiffusion modelFlow-based ModelAuto EncoderImage
🎯 What it does: Investigate the reconstruction-generation trade-off in high-dimensional latent spaces, revealing how encoders and decoders handle high-frequency information through frequency analysis, and propose the FreqWarm frequency preheating strategy to enhance the diffusibility of diffusion models in high-dimensional latent spaces.
Toward Early Quality Assessment of Text-to-Image Diffusion Models
Huanlei Guo (Southern University of Science and Technology), Bingyi Jing (Chinese University of Hong Kong)
GenerationConvolutional Neural NetworkVision Language ModelDiffusion modelContrastive LearningMultimodality
🎯 What it does: This paper proposes Probe-Select, a pluggable module that inserts lightweight detectors into early U-Net activations of diffusion models to enable early quality assessment of text-image generation, thereby terminating low-quality trajectories prematurely in a generate-select workflow.
Toward Generalizable Whole Brain Representations with High-Resolution Light-Sheet Data
Minyoung E. Kim (LifeCanvas Technologies), Brian Nguyen (LifeCanvas Technologies)
Object DetectionRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImageBiomedical DataBenchmark
🎯 What it does: Proposed the CANVAS high-resolution whole-brain LSFM dataset and benchmark, constructed a baseline detection model, and conducted self-supervised feature learning experiments.
Toward Low-Cost yet Effective Temporal Learning for UAV Tracking
Chaocan Xue (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
Object TrackingComputational EfficiencyTransformerVideo
🎯 What it does: Proposed a low-cost and efficient time learning method (LETL), integrated into the single-stream visual tracker LETrack based on DeiT-Tiny, to achieve real-time robustness for UAV visual tracking.
Toward Real-world Infrared Image Super-Resolution: A Unified Autoregressive Framework and Benchmark Dataset
Yang Zou (Northwestern Polytechnical University), Jinyuan Liu (Dalian University of Technology)
Super ResolutionAuto EncoderImageBenchmark
🎯 What it does: This paper proposes a unified autoregressive framework, Real-IISR, for real-world infrared image super-resolution, and constructs the FLIR-IISR dataset.
Towards an Incremental Unified Multimodal Anomaly Detection: Augmenting Multimodal Denoising From an Information Bottleneck Perspective
Kaifang Long (Northeastern University), Guoyang Xie (CATL)
Anomaly DetectionConvolutional Neural NetworkAuto EncoderMultimodality
🎯 What it does: Propose an incremental unified multimodal anomaly detection framework named IB-IUMAD, addressing catastrophic forgetting caused by pseudo feature interference and redundant information in cross-modal features.
Towards Balanced Multi-Modal Learning in 3D Human Pose Estimation
Mengshi Qi (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
Pose EstimationTransformerMultimodality
🎯 What it does: Proposed a balanced multi-modal learning framework that evaluates the contribution of each modality in 3D pose estimation using Shapley values, and balances the learning rates of different modalities early in training through adaptive weight constraints (AWC) based on the Fisher information matrix, ultimately achieving collaborative learning across four modalities: RGB, LiDAR, mmWave, and WiFi.
Towards Calibrating Prompt Tuning of Vision- Language Models
Ashshak Sharifdeen (Mohamed bin Zayed University of AI), Muhammad Haris Khan (Mohamed bin Zayed University of AI)
Explainability and InterpretabilityRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Significantly improve the model's confidence calibration ability on base and novel classes by incorporating mean-variance margin regularization and text moment matching regularization into the prompt-tuning process of CLIP.
Towards Cross-Modal Preservation, Consistency and Alignment for Privacy-Preserving Visible-Infrared Person Re-Identification
Yudi Xie (Wuhan University), Mang Ye (Wuhan University)
RecognitionPose EstimationRetrievalSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImageMultimodality
🎯 What it does: Propose a cross-modal anonymizable person re-identification framework named PPA, achieving re-identification of visible-infrared images while protecting privacy.
Towards Decompositional Human Motion Generation with Energy-Based Diffusion Models
Jianrong Zhang (University of Technology Sydney), Yi Yang (Zhejiang University)
GenerationDiffusion modelVideoTextSequential
🎯 What it does: Decompose human motion into composable motion concepts and achieve reusable motion synthesis based on an energy diffusion model.
Towards Dynamic Modality Alignment in Multimodal Continual Learning
Jiayao Tan (Tianjin University), Wei Feng (Tianjin University)
Representation LearningGraph Neural NetworkPrompt EngineeringContrastive LearningMultimodality
🎯 What it does: Proposes a dynamic modal alignment method for multi-modal continual learning - Dynamic Alignment Graph Regularization (DAGR), which stabilizes cross-layer alignment and suppresses catastrophic forgetting by constructing hierarchical dynamic alignment graphs in each task and applying multi-level graph regularization.
Towards Efficient Medical Reasoning with Minimal Fine-Tuning Data
Xinlin Zhuang (East China Normal University), Imran Razzak (MBZUAI)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: For medical reasoning tasks, we propose a data selection framework called DIQ (Difficulty-Influence Quadrant), which selects a subset that significantly improves model performance under small sample sizes by jointly measuring the reasoning difficulty of samples and gradient influence, thus achieving results comparable to full training with only a small amount of fine-tuning.